JP6414966B2 - Health data analysis auxiliary device - Google Patents

Description

  The present invention relates to a health data analysis auxiliary device that enables efficient health data acquisition.

  As pointed out in Non-Patent Document 1, the health management of employees (employees) is an important matter for corporate management. Therefore, many systems for managing health for employees and other fixed groups have been proposed. For example, Patent Document 1 proposes a mechanism for performing employee authentication and acquiring health data. On the other hand, wearable devices represented by FitBit (registered trademark) in Patent Document 2 and Non-Patent Document 1 have permeated, and using such simple devices, the health status of employees and health insurance members is acquired. can do.

  In addition, health checkups (hereinafter referred to as health checkups) are conducted nationwide, and health checkup data for all members is acquired once a year to extract high-risk individuals.

JP 2008-305050 A Patent No. 5489105

http://www.jmsc.co.jp/word/column/no83.html http://www.fitbit.com/jp

  However, since general health examinations have limited health examination items, it is possible to extract very high-risk people who have already developed a disease or those who have already developed disease, but most potential Often overlook high-risk people. That is, there is a problem that it is difficult to extract a person who can be prevented even if only a person who is already difficult to prevent can be extracted. In other words, there are some people who can be adequately screened using only medical examination data, and there are some who cannot, and there are many high-risk people who can potentially be prevented among those who cannot.

  As a method of supplementing such data, for example, it is possible to increase the number of medical examination items or distribute wearable devices, but on the other hand, generally grasp the health status of employees in tens of thousands of large companies. In order to do this, it is not practical to distribute wearable devices. In addition, as for the medical examination items, if the amount of all employees is increased, the cost will be greatly increased.

  An object of the present invention is to provide a health data analysis auxiliary device that enables efficient health data acquisition in view of the above-described problems of the prior art.

  In order to achieve the above object, the present invention provides a first acquisition unit that acquires an evaluation value of each item of health data consisting of a first item group for a plurality of subjects, each item of the first item group, and health data A relationship matrix storage unit that stores a predetermined numerical value representing a relationship between each item of the second item group and a relationship matrix, and the evaluation value of each item of the acquired first item group, the relationship In the sex matrix, a value given as a weight for each item of the corresponding second item group is generated as a feature vector of each subject specialized for each item of the second item group, and the second item group For each item, a clustering unit that clusters the plurality of subjects based on the feature vector, and each cluster of the result of clustering for each item of the second item group, each item of the second item group VS to evaluate , A representative determination unit that determines a representative as characterized in that it comprises a.

  According to the present invention, the health data consisting of the first item group, and the relationship matrix that quantifies the relationship between each item of the first item group and each item of the second item group of the health data are input. As described above, a plurality of subjects can be clustered in a form specialized for each item of the second item group, and a representative as an object for evaluating each item of the two item group can be determined from each cluster. Therefore, effective screening when actually evaluating the health data of the second item group without using the health data of the second item group as an input will be achieved by the determined representative. Health data can be acquired.

  In particular, when the health data of the first item group is easily obtainable and the health data of the second item group is difficult to obtain, effective acquisition of the health data of the second item group is possible. It becomes possible.

It is a functional block diagram of the health data analysis auxiliary device concerning one embodiment. It is a flowchart of operation | movement of the health data analysis assistance apparatus which concerns on one Embodiment. It is a figure which shows the example of the item in medical examination data which can be utilized as health data of a 1st item group. It is a figure which shows the example of the item in the health data acquired with the wearable device which can be utilized as health data of a 2nd item group. It is a figure which shows the typical example of the value of a relationship matrix. It is a figure which shows the conceptual example of a cross tabulation table. It is a figure for demonstrating the clustering by a latent topic analysis. It is a figure which shows typically the classification result of a clustering part. It is a figure for demonstrating the specific calculation when a reanalysis judgment part corrects a relationship matrix using an example. It is a figure which shows the example of the relationship matrix of the person who can lend mutually. It is a figure which shows the example of a relationship matrix and the result of the Markov clustering.

  FIG. 1 is a functional block diagram of a health data analysis auxiliary device according to an embodiment. The health data analysis auxiliary device 10 includes a first acquisition unit 1, a clustering unit 2, a relationship matrix storage unit 3, a representative determination unit 4, a second acquisition unit 5, an estimation unit 6, a reanalysis determination unit 7, and a first quantification A unit 11 and a second digitizing unit 15.

  FIG. 2 is a flowchart of the operation of the health data analysis apparatus according to the embodiment. Hereinafter, each part of FIG. 1 will be described while explaining each step of FIG. 2.

In step S1, the first acquisition unit 1 acquires health data of the first item group for a plurality of subjects. Let each of the first item group be E1 (i) (i = 1, 2,..., N). Each item E1 (i) is an item for evaluating health, and the evaluation value is given to each target person in the acquired health data. If the evaluation value of item E1 (i) of any one subject a among a plurality of subjects is e1 _{[i, a]} , the first obtaining unit 1 uses the obtained health data to obtain a plurality of subjects. Thus, a vector (e1 _{[1, a]} , e1 _{[2, a]} ,..., E1 _{[n, a]} ) representing the health state of each of the subjects a is obtained.

  Specifically, for example, health check data can be used as the health data of the first item group acquired in step S1. In addition, since the medical examination is performed once a year and the implementation rate is high, the medical examination data can be easily acquired for a large number of subjects.

  FIG. 3 shows an example of items of medical examination data. Here, the items in the medical examination data are roughly divided into three, and the items related to “Body data”, “Blood data” and “Questionnaire” are shown in [1], [2], [3], respectively. ing.

In step S1, the first acquisition unit 1 acquires the health data of the first item group and represents the health state of each subject a (e1 _{[1, a]} , e1 _{[2, a]} , ..., When obtaining e1 _{[n, a]} ), it is necessary to obtain the vector in a format that can be clustered in step S2 described later. The first digitizing unit 11 is responsible for the conversion of the format in step S1.

  That is, as the health data of the first item group, the medical examination data as shown in FIG. 3 can be used, but generally, the evaluation value of each item E1 (i) (i = 1, 2,..., N) is quantitative data. In some cases, it may be given as qualitative data. For example, the evaluation value of the item “hope for health guidance” in the questionnaire [3] in FIG. 3 is “Yes” or “No” as qualitative data. Moreover, even if it is quantitative data, the unit changes with items. For example, in the body data item of [1] in FIG. 3, the unit of “height” is “length”, and the unit of “weight” is “weight”.

Therefore, in the “raw data” state, the evaluation values of each item E1 (i) (i = 1, 2,…, n) take various forms as described above. By converting to the format, the above-mentioned vector (e1 _{[1, a]} , e1 _{[2, a]} ,…, e1 _{[n, a]} ) representing the health status of each subject a that can be clustered is generated. Is the first numerical unit 11.

  For the conversion by the first digitizing unit 11, a predetermined conversion formula is prepared in advance for each item E1 (i) (i = 1, 2,..., N), and the conversion formula is applied. Alternatively, Japanese Patent Application No. 2013-159323 (numerical data analysis apparatus and program), Japanese Patent Application No. 2013-163207 (numerical data analysis apparatus and program), Japanese Patent Application No. 2013-217817 (numerical data documentation) Device and program) may be used. If the health data of the first item group is preliminarily given in the unified numerical format as described above, the conversion process by the first digitizing unit 11 is omitted. You can do it.

In step S2, the clustering unit 2 clusters a plurality of subjects constituting the health data of the first item group acquired by the first acquisition unit 1 in step S1. Here, a vector (e1 _{[1, a]} , e1 _{[2, a]} ,…, e1 _{[n, a]} ) (= V (a)) (= V (a)) representing the health status of each subject a Clustering is performed by using the feature vector V (a) representing a.

However, in this case, V (a) is not clustered as it is, but the degree of influence of each item E2 (i) (i = 1, 2,…, m) of the health data of the second item group is represented by the feature vector V (a ) In advance, the feature vector V (a) _{[E2 (i)]} specialized for the item E2 (i) of the subject a is converted into the feature vector V (a) _{[E2 (i) ]} clustering is implemented as target.

  Accordingly, a clustering result is obtained for each item E2 (i) (i = 1, 2,..., M) of the second item group. Each cluster of the clustering result for the item E2 (i) is denoted as C (i, k) (k = 1, 2,..., K (i)). K (i) is the number of clusters of the i-th item E2 (i). Here, the number of clusters may be arbitrary (predetermined number) for each item E2 (i), an optimal number may be determined by AIC (Akaike Information Criterion), or an infinite dimensional LDA or mixed dirichlet. You may make it determine automatically by using a process etc.

  Specifically, each item of the health data of the second item group can use, for example, health data acquired by various wearable devices. FIG. 4 shows an example of items of health data acquired by the wearable device.

  Here, the health data of the second item group is more difficult to obtain than the health data of the first item group, but instead, the health data gives more detailed information. Assumed. For example, the medical examination data as an example of the health data of the first item group can be easily acquired for a large number of subjects because the medical examination is widely performed regularly, but only a rough health condition can be grasped. It is assumed that it is a thing.

  On the other hand, the health data of the wearable device as an example of the health data of the second item group is difficult to distribute the wearable device, which is a special device, to a large number of subjects similar to the medical examination. In addition, although it is difficult to acquire the data due to the trouble of actually acquiring the data by wearing the wearable device, the detailed information on the health condition as shown in FIG. 4 is given instead. It is assumed that it is a thing.

  Therefore, for example, when the health examination data as the health data of the first item group is that of the “standard” health examination, the health data of the second item group is changed to the standard health examination (from the viewpoint of cost burden). It is good also as a thing of the "special" medical examination implemented about the detailed medical examination item abbreviate | omitted.

Feature vector V (a) specialized for each item E2 (i) of the second item group of feature vector V (a) representing target person a described above as preprocessing of clustering by clustering unit 2 in step S2 _[ For conversion to _{E2 (i)]} , if the relationship matrix stored in advance in the relationship matrix storage unit 3 is referred to, each element of the matrix is added to each element of the vector as a weight. Good.

The relationship matrix M is a numerical value representing the relationship between the value M _{i, j} of the i row and j column element of the item E2 (i) in the second item group and the item E1 (j) in the first item group ( The larger the relationship, the larger the value, and the weaker the relationship, the smaller the value. (Although it may be defined by this transposition, since the meaning is the same, hereinafter, the relationship matrix M is referred to in the definition.)

By using the relationship matrix M = (M _ij ), the feature vector V (a) = (e1 _{[1, a]} , e1 _{[2, a]} ,…, e1 _{[n, a]} ) Feature vector V (a) _{[E2 (i)]} specialized for each item E2 (i) in the item group = (M _{i, 1} * e1 _{[1, a]} , M _{i, 2} * e1 _{[2, a]} ,…, M _{i, n} * e1 _{[n, a]} ). That is, in the feature vector V (a), the evaluation value “e1 _{[j, a]} ” of each item E1 (j) (j = 1, 2,..., N) of the first item group is listed as an element. Therefore, for each of these, the weighted evaluation value “M _i is obtained by integrating the values M _{i, j} of the relationship between the item E2 (i) of the second item group and the item E1 (j) as weights. _{, j} * e1 _{[j, a]} ”to generate a feature vector V (a) _{[E2 (i)]} .

  FIG. 5 shows a schematic example of the value of the relationship matrix. In the example of FIG. 5, an example of values of a part of the relationship matrix is shown as a table, items E2 (1) and E2 (2) in the second item group in the row direction, and items in the first item group in the column direction. E1 (1), E1 (2), E1 (3), E1 (4), and E1 (5) are set with items as shown in the figure as specific examples. For example, the numerical value of the relationship between item E2 (2) `` activity amount '' in the second item group and item E1 (5) `` exercise habits of 30 minutes or more '' in the first item group is the two rows in the table. By referring to the position of 5 columns (position excluding the item display part), it can be seen that it is “0.9”.

  The relationship matrix M may be calculated in advance and stored in the relationship matrix storage unit 3 and referenced from the clustering unit 2. Specific methods for calculating the relationship matrix M include [Method 1] to [Method 3] below.

[Method 1] By using general knowledge, a predetermined value is set for each element of the relationship matrix M = (M _{i, j} ).

[Method 2] Aggregating the data of people who have both data of item E2 (i) and item E1 (j), obtaining correlation by well-known multivariate analysis etc., and using the correlation value, the relationship matrix M = _Let (M _{i, j} ).

  In [Method 2], a well-known cross tabulation table independence test may be performed instead of the multivariate analysis. FIG. 6 shows a conceptual example of a cross tabulation table regarding “step count” as an example of item E2 (i) and “BMI” (body mass index) as an example of item E1 (j). Specific numerical values in each column are omitted). In this case, for example, “1” may be set for items determined to be related, and “0” may be set for non-related items, or a statistic may be calculated and used as a function of the statistic. It may be set continuously.

  [Method 3] Obtain a web document on the Internet, analyze the document by applying a well-known text analysis technique, and calculate the relationship from the co-occurrence and distance between words. As the word, the item E2 (i) and the item E1 (j) may be used as they are, or related words may be used.

As described above, in Step S2, the clustering unit 2 assigns the feature matrix V (a) representing each subject a to each item E2 (i) of the second item group by assigning the relationship matrix M _{i, j} as a weight. after converted into a feature vector V specialized _{(a) [E2 (i)} ] to, specialized in the feature vector V _{(a) [E2 (i)} ] item E2 (i) by targeting Perform clustering. For the clustering, a known method such as LDA (latent Dirichlet allocation method) which is a kind of k-means or latent topic analysis can be used.

  Here, since it is a premise in the description of the items described later, clustering by latent topic analysis such as LDA will be briefly described. FIG. 7 is a diagram for explaining the result of clustering by the latent topic analysis and its meaning.

In FIG. 7, all data to be clustered obtained from the first acquisition unit 1 is D, and data in a matrix format in which the feature vectors V (a) _{[E2 (i)]} of each subject a are listed as row vectors It is. In latent topic analysis, each row of the matrix D is a "document u" as Baguobuwazu, elements aligned in the column direction each word i _1, i _2, ..., is the frequency number of the i _n.

In the present invention, each “document u” corresponds to each “subject a”, and each word i ₁ , i ₂ ,..., I _n corresponds to each item E1 (1), E1 (2 ), ..., E1 (n). Then, the specific frequency numbers of the respective words i ₁ , i ₂ ,..., I _n are values as calculated as the feature vector V (a) _{[E2 (i)]} .

  In latent topic analysis such as LDA, the document u (each target person a in the present invention) is treated as a bug of words, that is, a word and its appearance frequency, and the topic is estimated in the document. For example, words such as “stock price”, “increased sales and profits” will appear from the “economy” topic, and words such as “baseball” and “soccer” will appear from the “sports” topic.

  This is because the observed bug-of-words expression, that is, the relationship matrix D between the word i (column component) and the document u (row component), as shown in FIG. 7, the word i (column component) and the topic k (row component). , And the product of the relationship matrix θ of the document u (row component) and the topic k (column component). This topic k is estimated by latent topic analysis such as LDA, and as a result of the analysis, a decomposition result of “D = θ × Φ” as shown in FIG. 7 is obtained.

  Thus, the topic model represented by LDA assumes that each document has a unique topic ratio, and words are generated at a ratio specific to the topic after selecting the topic according to this topic ratio. .

  When interpreting the clustering result of each document u (= target person a), for example, the row (row vector) of the document u in the relationship matrix θ represents the affiliation probability of the document u to each topic k. Therefore, for example, it can be interpreted as belonging to the cluster corresponding to the topic k having the highest membership probability.

  Returning to the description of each step in FIG. 2, in step S3, the representative determining unit 4 classifies the clustering unit 2 for each item E2 (i) (i = 1, 2,..., M) of the second item group. From each of the clusters C (i, k) (k = 1, 2,..., K (i)), a representative to be actually evaluated for the item E2 (i) is determined.

  The significance of the representative to be determined in the present invention is as follows.

  Particularly in the present invention, the classification result of the clustering unit 2 is classified by paying attention to the item E2 (i) in the health state in particular by assigning the relationship matrix M in which appropriate values are set in advance as weights. It is the result. That is, the evaluation results of the item E2 (i) are generally similar between the subjects belonging to each cluster C (i, k). Therefore, by evaluating the item E2 (i) only for a small number of representatives determined from each cluster C (i, k), all items belonging to the cluster C (i, k) with respect to the item E2 (i) You can know the tendency of the target audience.

  For this reason, even if the item E2 (i) of the second item group is acquired by the wearable device as described in FIG. 4 and is difficult to acquire from the viewpoint of cost or the like, For all subjects for whom health data for a group of items was acquired, item E2 () was only available for representatives of each cluster C (i, k) without incurring high costs and evaluating item E2 (i). By evaluating i), it becomes possible to know the tendency of item E2 (i) of all subjects.

  Therefore, if the result of investigation by the representative of cluster C (i, k) shows a dangerous (or potentially dangerous) result for item E2 (i), all members in cluster C (i, k) Is likely to be a dangerous (or potentially dangerous) result for item E2 (i), so item E2 (i) will be evaluated for each subject belonging to cluster C (i, k). You just have to do it.

  That is, according to the present invention, with respect to the item E2 (i), which is a high-cost evaluation item, only the representatives of the cluster C (i, k) are preliminarily evaluated, so that the survey is focused on among all the subjects. It becomes possible to appropriately screen (select) the subjects to be used. As a result, it is possible to find a potentially preventable high-risk person who has been overlooked only by conventional normal medical examination data at a low cost.

  FIG. 8 is a diagram schematically illustrating the classification result of the clustering unit 2, and each cluster includes a target a as an example of a part of all the target persons from which health data of the first item group is acquired. , b, c, d, e are drawn. In [1], as an example of the item E2 (1), a clustering result specialized for “step count” is shown as C (1,1), C (1,2),. In [2], as an example of item E2 (2), the clustering result specialized for “heartbeat” is shown as C (2,1), C (2,2),. In [3], as an example of the item E2 (3), the clustering result specialized in “activity” is shown as C (3,1), C (3,2),.

  On the premise that clustering was performed using an appropriate relationship matrix M, for example, a common evaluation regarding the “step count” between subjects a and b belonging to cluster C (1,1) in [1] As a result, a substantially common evaluation value regarding the “step count” can be obtained among the subjects c, d, e belonging to the cluster C (1, 2). Further, different evaluation values related to the “number of steps” between the target persons are obtained between different clusters C (1,1) and C (1,2).

  In order for the representative determining unit 4 to specifically determine the representative having the above significance in step S3, first, it is necessary to determine the number of representatives for each cluster. The following [Method 11] or [Method 12] can be used.

  The clustering result C (i, k) (k = 1, 2,…, K (i)) is obtained for each item E2 (i) (i = 1, 2,…, m) of the second item group. Similarly to the existence, the total number N (i) of representatives determined for each item E2 (i) is given a predetermined value. The total number N (i) may be set in advance in consideration of the cost that can be paid when the item E2 (i) is actually evaluated (budget relationship, etc.).

  [Method 11] The number of representatives of each cluster C (i, k) is determined so as to be proportional to the number corresponding to the variation (variance) of the evaluation values of the subjects belonging to each cluster C (i, k). Here, the larger the variation, the more the health characteristics of the subjects in the cluster are not constant (that is, an inaccurate cluster), so a predetermined monotonically increasing function is applied to the variation. It is preferable that the number is determined so as to be proportional to the number determined.

At this time, for the evaluation value as a target for calculating variation, the above-described feature vector V (a) _{[E2 (i)]} for each target person a may be used, or clustering is performed by latent topic analysis such as LDA. If so, the topic affiliation probability of each target person a may be used. The topic affiliation probability is given as each row vector in the relationship matrix θ between the document u (= target person a) and the topic k in FIG.

  [Method 12] The number of representatives of each cluster C (i, k) is determined to be proportional to the number of subjects belonging to the cluster.

  After determining the number of representatives of each cluster C (i, k) as described above, the following [Method 13] or [Method 14] can be used.

  [Method 13] Random selection using random numbers or the like.

  [Method 14] In advance, from the target person (applicant) who wishes to be subject to the evaluation of the item E2 (i), information on the request is received. Is first determined as a representative, and the remaining target persons are determined by the above [Method 13]. In addition, when the applicant has reached the number of representatives of the cluster, the determination may be made based on a predetermined criterion such as giving priority to a person who has received the information indicating that the request is desired.

  Further, when the representative is determined as described above, the following [Constraint 1] or [Constraint 2] may be imposed. Which restrictions are imposed depends on the operation policy and the like.

  [Restriction 1] A subject who is determined as a representative of a cluster C (i, k) with a certain predetermined item E2 (i) is not determined as a representative of a cluster with another predetermined item E2 (j) Or lower the priority to be determined.

  [Restriction 2] Contrary to the above, the target person determined as the representative of a certain cluster C (i, k) of a certain predetermined item E2 (i) is the target of the cluster of another predetermined item E2 (j). Decide as a representative or raise the priority of the decision.

As described above, when the representative determination unit 4 determines a representative for the clustering result of each item E2 (i) in step S3, the representative determination unit 4 notifies the corresponding representative of the determination result, thereby notifying. In each of the representatives who received the evaluation, the evaluation of the item E2 (i) is actually performed, and in step S4, the second acquisition unit 5 evaluates the health data of the series of the second item group obtained by the execution. Get the value. The evaluation value of the first item group is expressed in the same notation as e1 _{[i, a],} and the evaluation value of the item E2 (i) in the second item group for the representative _r is expressed as e2 _{[i, r]} . .

  Note that, when each evaluation value e2 (i) is acquired in step S4, it may be converted into a numerical value by the second numerical conversion unit 15 and then acquired by the second acquisition unit 5. The function of the second digitizing unit 15 for digitizing is the same as that of the first digitizing unit 11.

  Further, the notification to the representative in step S3 may be performed by a well-known e-mail or the like. It is to be noted that the actual evaluation of item E2 (i) in each of the representatives who received the notification is an artificial work outside the health data analysis auxiliary device 10, and the work is performed in steps S3 and S4. Between. The same applies to the evaluation of the health data of the first item group acquired by the first acquisition unit 1 (execution of medical examination, etc.).

  Steps S4 and subsequent steps are steps related to operations and the like when actually performing screening and the like by the representative determined in step S3. A user who uses the health data analysis assisting device 10 (assuming an analyst of health data) Can obtain various knowledge about health analysis in each step. In the following, typical operations will be described, but various operations can be performed according to the operation policy.

  In step S5, the use of the health data of the series of second item groups acquired by the second acquisition unit 5 in step S4 is performed. In other words, the use is to extract various useful information from the health data of the second item group. In one embodiment, the estimation unit 6 performs estimation, and in one embodiment, the reanalysis determination unit 7 performs reanalysis. It is possible to perform a determination as to whether or not to perform a reanalysis if a determination is made. Hereinafter, each embodiment by the respective units 6 and 7 will be described.

The estimation unit 6 determines the evaluation value e2 _{(i, a} ) of the item E2 (i) regarding the subject a for which the health data of the item E2 (i) of the second item group has not been acquired as _{a representative. ]} Is estimated. In the estimation, evaluation values e2 _{[i, r]} in a series of representatives r from which the health data of E2 (i) are actually acquired are used.

If the clustering unit 2 performs clustering on the premise of a single topic such as k-means, it is necessary that the target person a and the representative r to be estimated belong to the same cluster. As “e2 _{[i, a]} = e2 _{[i, r]} ”. If there are a plurality of representatives r in the cluster, the evaluation value e2 _{[i, a]} of the target person a who is not the representative in the cluster is the evaluation value e2 _[i, It can be estimated as given by a statistical distribution similar to _r] . For example, the evaluation value e2 _{[i, a]} may be estimated as an average value in the distribution.

  On the other hand, if clustering unit 2 performs clustering on the premise of multiple topics such as LDA and other latent topic analysis, in addition to estimation by the same method as in the above k-means etc., each cluster The value may be estimated by the average value of the representatives and the weighted sum of the membership probabilities.

  That is, for the clustering result C (i, k) (k = 1, 2,..., K (i)) of an item E2 (i), the item E2 (i in the representative in the cluster C (i, k) ) Is the average of the evaluation values of e2m (i, k). In addition, the affiliation probability of the target person a to be estimated to each topic k (= 1, 2,..., K (i)) is assumed to be w (a, k). Note that, as described above, the affiliation probability w (a, k) is the row corresponding to the target person a in the relation matrix θ between the document u (= target person a) and the topic k described in FIG. The value of the element at the position specified in the column corresponding to topic k. (Note that the relationship matrix θ needs to be normalized in advance so that the affiliation probability w (a, k) of each subject person a is summed to 1 for all topics k.)

As described above, the estimated value e2 _{[i, a]} of the evaluation value of the item E2 (i) of the subject person a can be obtained by the following equation (1).

  Note that the affiliation probability w (a, k) is the affiliation probability of each subject a belonging to each topic k, that is, each cluster C (i, k).

The reanalysis determination unit 7 first determines whether or not the clustering result in the clustering unit 2 is appropriate as a specialized item E2 (i). For the determination, first, in each cluster C (i, k) (k = 1, 2,..., K (i)) of the clustering result, items acquired from the representative r determined by the representative determination unit 4 When the average m (i, k) (k = 1, 2,…, K (i)) of the evaluation value e2 _{[i, r]} of E2 (i) is calculated, and the variance of the average is less than or equal to a predetermined threshold It is judged that it was inappropriate. (Conversely, if it exceeds a predetermined threshold value, it is determined to be appropriate.)

  Here, when the variance of the average m (i, k) is less than or equal to the predetermined threshold, there is no particular difference in the representative evaluation values between the different clusters C (i, k1) and C (i, k2). As a result, it is determined that the clustering result was inappropriate.

  Furthermore, when the re-analysis determination unit 7 determines that the clustering result of the item E2 (i) is inappropriate, the re-analysis determination unit 7 selects the inappropriate clustering result from the initial relationship matrix stored in the relationship matrix storage unit 3. Modify the value of each column of i row that was given.

  FIG. 9 is a diagram for explaining a specific calculation when the correction is performed using an example.

  In FIG. 9, [1] indicates that the Φ matrix (as described in FIG. 6) generated when the clustering unit 2 performs clustering specialized for “sleep” as the item E2 (i), for example. A relationship matrix between the topic k and the word i), and each cluster 1 of the evaluation value of the item E2 (i) actually acquired by the second acquisition unit 5 from the representative determined by the representative determination unit 4 according to this , 2 and 3 mean and variance. (The variance is a standardized variance.)

  As described above, the “word” corresponds to each item E1 (i) of the first item group, but “word” is used as a term in potential topic analysis such as LDA. In the example of FIG. 9, “BMI”, “abdominal circumference”, and “HbA1c” are shown as three “words”.

  In the example [1], since there is almost no difference in the average value between the clusters 1, 2, and 3, it is determined that the clustering result is inappropriate, and the relationship matrix is corrected. Here, the “concept” at the time of the correction is as follows. That is, it is considered that a cluster with a large variance is not able to perform the original clustering well, so the weight in the relationship matrix of words that contribute to the generation of such a cluster should be reduced.

  Based on the above “concept”, as a specific calculation, in each cluster, the variance (standardized variance) may be weighted and multiplied by the word occurrence probability of each Φ matrix.

  [2] in FIG. 9 shows an example of multiplication of the weight in each cluster 1, 2, and 3 and the word occurrence probability of each Φ matrix. The upper part of [3] shows an example of calculating the weight for each word by adding the weights of the clusters 1, 2, and 3 for each word. The lower part of [3] further shows an example of calculation for normalizing the weight for each word.

In this way, the weight W _{i, j} normalized for each item E1 (j) in the item E2 (i) (= words “BMI”, “abdomen”, “HbA1c” in FIG. 9) can be obtained. When correcting the relationship matrix M = (M _{i, j} ), the value of the corresponding element may be reduced by the calculated weight. For example, M _{i, j} may be corrected to M ′ _{i, j} by the following equations (2) and (3). In Expression (3), c is a constant and is determined within a range such that M ′ _{i, j} > 0.
M ' _{i, j} = M _{i, j} / W _{i, j} (2)
M ' _{i, j} = M _{i, j} -c * W _{i, j} (3)

  When the relationship matrix M is corrected as described above, the corrected relationship matrix M ′ is updated and stored in the relationship matrix storage unit 3, and then the corrected item E2 (i) Again, step S2 and subsequent steps in FIG. 2 can be performed. If the re-executed result is inappropriate, it can be repeated in the same manner until an appropriate result is obtained.

  As an application of repeating until the appropriate result is obtained, the following is also possible. In other words, the relationship matrix stored in the relationship matrix storage unit 3 is initially set to a random value, or all elements are set to 1, and the relationship matrix is repeated without prior knowledge by repeating until an appropriate result is obtained. Can be automatically generated. As described above, [Method 1] to [Method 3] have been described as methods for calculating the relationship matrix, but this method can also be used as [Method 4].

  Here, an embodiment will be described in which restrictions on actual operation in a company or the like can be further considered in representative determination by the representative determination unit 4 in step S3. As a specific situation, each item E2 (i) of the second item group is acquired by a wearable device, and the evaluation value is acquired by lending the wearable device to a representative. . As a constraint at this time, it is assumed that there are fewer wearable devices that can actually be lent than the total number of representatives whose evaluation values are to be acquired.

  In such a case, there is no choice but to lend in sequence at intervals. Therefore, the representative determining unit 4 may determine and notify not only the representative but also the lending schedule.

  Here, in one embodiment, the lending period (the total period during which lending is allowed), the period necessary for data acquisition, and information on persons who can lend each other are input to the representative determination unit 4 as prior knowledge. The lending order (schedule) can be automatically determined. (This person who can lend to each other can also be determined from the department information if it is a corporate organization. That is, it is determined that the device can be rented in the same organization or in an adjacent organization, and otherwise. (Determining that the loan is impossible)

Hereinafter, an example in which the lending order is automatically determined will be described. First, the prior knowledge to be entered is as follows.
・ Determined representative: a, b, c
・ Lending period: 2 weeks ・ Necessary period for data acquisition: 1 week ・ Relationship matrix of people who can lend each other between a, b, and c (Figure 10)

  Here, in FIG. 10, for the representatives a, b, and c, direct lending is possible, for example, because the departments to which they belong are close to each other (delivery of the device from the representative who has finished the measurement to the next representative to be measured) In the case of lending without generating a time lag, the relationship matrix is given by setting 1 to 0, otherwise 0, and since the diagonal component is meaningless to 0 .

  At this time, only paths that can be reached within the steps of lending period / period necessary for data acquisition = 2 steps among the paths that can be reached from the relationship matrix of FIG. 10 are allowed. The following is a list of reachable paths for each number of steps.

・ 1 step
a, b, c
・ 2 steps
b → a
b → c
・ 3 steps
a → b → c
c → b → a

  Of these, up to two steps are allowed, so when lending a specific device to a, b, c, prepare two devices and lend b → a and c, or b → c. You can rent a.

  That is, in this embodiment, under the constraints of the lending period, the period required for data acquisition, and information on persons who can lend each other, the number of devices and lending order required to satisfy the restrictions are determined. Can be determined automatically.

  On the other hand, in another embodiment, it is possible depending on the progress of the actual lending, with restrictions on the information of people who can lend each other and the number of devices (the number of devices is a predetermined number less than the number of representatives). It is also possible to automatically determine the lending order for completing data acquisition in as short a period as possible. In this case, the period required for data acquisition may be fixed, or may vary for each representative for whom lending has been performed.

  Specifically, in the embodiment, after obtaining the relationship between representatives in the following [Precalculation 1] and [Precalculation 2], the lending order (in accordance with [Policy 1] to [Policy 3] automatically ( Lending schedule).

[Preliminary calculation 1] Markov clustering is performed on the relationship matrix that represents the information of people who can lend each other. Here, Markov clustering is well known as introduced in the following URL and the like. With Markov clustering, nodes that are easy to move to each other (nodes with dense edges) can be grouped in a graph in which edges are provided between people who can lend each other as nodes in the relationship matrix. . That is, in device lending, it is possible to cluster together people who are easy to lend each other.
(URL) http://rishida.hatenablog.com/entry/2013/09/24/011423

  FIG. 11 shows an example of a relationship matrix (an example different from that of FIG. 10) and the result of the Markov clustering. [1] shows an example of a relationship matrix for seven people A to G, and the corresponding graph is shown in [2]. As the results of the Markov clustering of [1] and [2] are shown in [3] and [4], respectively, the cluster consisting of A, B, C, D and the cluster consisting of E, F, G It is divided.

[Pre-calculation 2] For each node (that is, each person) of the graph represented by the relationship matrix, a well-known page rank is calculated with the following URL or the like.
(URL) http://deta.hateblo.jp/entry/2013/05/19/060609

For example, the page rank value is calculated for each of the nodes A to G in FIG. B is higher in the cluster composed of A, B, C, and D, and E is larger in the cluster composed of E, F, and G because of the higher connection rank than other nodes.
[A] 0.09060 [B] 0.10125 [C] 0.09060 [D] 0.09060 [E] 0.09770 [F] 0.08918 [G] 0.08918

  [Policy 1] Lending is performed for each cluster obtained in [Preliminary calculation 1]. The number of lending devices in each cluster may be, for example, the total number of devices divided by the ratio of the number of people belonging to the cluster.

  [Policy 2] For lending within each cluster, first lend devices to people with low page rank. In other words, if the cluster is composed of N people and n devices (n <N) are allocated for lending, the first lending target is the page rank of the lowest N people from Nth place to N-1 place, N-2 place, ... N-n + 1 place.

  [Policy 3] Regarding the complete person X who has received the device lending and has completed the measurement by himself, it is necessary to determine the target person Y who is to lend the device in the cluster next. Regarding the decision, based on the time (date and time) when the measurement was completed by the completed person X, the decision was automatically made from those who satisfy the following (Condition 1) and (Condition 2), and the measurement was completed. It is only necessary to automatically notify the completion person X and the next rental target person Y so that the device is delivered.

(Condition 1) The target person Y is in a relationship that allows immediate device transfer with the complete person X, that is, the value representing the relationship between the target person Y and the complete person X in the relationship matrix is 1.
(Condition 2) Target Y has not yet rented a device.
(Decision policy) Among the series of target persons Y that satisfy the above (Condition 1) and (Condition 2), the page rank value PR (Y) is greater than or equal to the page rank value PR (X) of the complete person X A person (PR (Y) ≧ PR (X)) whose difference PR (Y) -PR (X) is the smallest is determined as the target person Y who should actually lend the device next. .

  As described above, by continually determining the next lending target according to [Policy 3] until the measurement is completed after lending a device to all the target persons in the cluster, the cluster can be efficiently used as follows. It is possible to lend and measure to everyone in the company. In other words, lending is given priority to nodes with low page rank (the number of nodes that can lend each other is low), and delays associated with device delivery are possible by postponing nodes with high page rank that are easy to lend to each other. Lending to everyone can be done efficiently while avoiding as much as possible.

  Note that when [Measure 3] is unable to determine the next Target Y when measurement has not been completed for all subjects in the cluster, [Policy 3-1] will be used instead of [Policy 3]. ] Or [Policy 3-2]. [Policy 3-1] continues to lend out from the low page rank value to the high one, and [Policy 3-2] reverses the policy and reverses the page rank value from the high Lending to the side will continue. If it is impossible to determine the next target person in [Policy 3-2], the process returns to [Policy 3-1] again. After that, [Policy 3-1] and [Policy 3] -2] may be switched.

  [Policy 3-1] Among the target persons Z before the device rental in the cluster, the target person Z having the smallest page rank value PR (Z) may be set as the next rental target person. In this case, if there is a subject Z that satisfies (Condition 1), the subject Z that has the smallest page rank value PR (Z) among the subject Z that satisfies (Condition 1) is It is assumed to be lent out. The policy 3 will be re-applied for the target person Z and later.

  [Policy 3-2] Among the target persons Z before the device rental in the cluster, the target person Z having the maximum page rank value PR (Z) is set as the next rental target person. In this case, if there is a subject Z that satisfies (Condition 1), the subject Z that has the highest page rank value PR (Z) among the subject Z that satisfies (Condition 1) is It is good also as a loan object person. For the target person Z and later, [Policy 3] will be tried again, but (Decision Policy) will be changed to the following (Decision Policy 2).

  (Decision policy 2) Among the series of target persons Y that satisfy the above (Condition 1) and (Condition 2), the page rank value PR (Y) is less than or equal to the page rank value PR (X) of the complete person X (PR (Y) ≤ PR (X)), the person whose difference PR (X)-PR (Y) is the smallest is determined as the target person Y who should actually lend the device next To do.

  As described above, according to the present invention, based on the health data of each item of the first item group that can be easily acquired, the health data of each item of the second item group that is difficult to acquire but can obtain precise information is obtained. When obtaining, effective screening becomes possible.

  The present invention can also be provided as a program that causes a computer to function as the health data analysis auxiliary device 10. The computer can be configured with well-known hardware such as a CPU (Central Processing Unit), a memory, and various I / Fs, and the CPU that reads and executes the program functions as each part of the health data analysis auxiliary device 10. It becomes.

  10 ... health data analysis auxiliary device, 1 ... first acquisition unit, 2 ... clustering unit, 3 ... relationship matrix storage unit, 4 ... representative determination unit, 5 ... second acquisition unit, 6 ... estimation unit, 7 ... reanalysis Judgment unit, 11 ... first numerical unit, 15 ... second numerical unit

Claims (13)

A first acquisition unit that acquires an evaluation value of each item of health data consisting of a first item group for a plurality of subjects;
A relationship matrix storage unit that stores a predetermined numerical value representing a relationship between each item of the first item group and each item of the second item group of health data as a relationship matrix;
A value obtained by assigning a numerical value of each item of the corresponding second item group in the relationship matrix as a weight to the evaluation value of each item of the acquired first item group is specified for each item of the second item group. A clustering unit that generates the feature vectors of each target person and clusters the plurality of target persons based on the feature vectors for each item of the second item group,
A representative determining unit that determines a representative as a target for evaluating each item of the second item group from each cluster obtained by clustering each item of the second item group. Data analysis auxiliary device.   The health data analysis auxiliary device according to claim 1, wherein the representative determining unit determines the number of representatives of each cluster according to variations in the feature vectors of subjects belonging to each cluster.   The health data analysis auxiliary device according to claim 1, wherein the clustering unit performs the clustering by latent topic analysis.   The health data analysis auxiliary device according to claim 1, further comprising a second acquisition unit that acquires an evaluation value of each item of the second item group in the determined representative.   Based on the evaluation value of each item of the second item group in the acquired representative, an estimation unit that estimates the evaluation value of each item of the second item group in the subject not determined as the representative The health data analysis auxiliary device according to claim 4, further comprising: The clustering unit obtains a probability of belonging to each cluster of each subject as a clustering result by performing the clustering by latent topic analysis,
The estimation unit is not determined as the representative as a weighted sum based on an affiliation probability in each target person with respect to an average evaluation value of each item of the second item group of the acquired representative in each cluster. The health data analysis auxiliary device according to claim 5, wherein an evaluation value of each item of the second item group in the subject is estimated.   The health data analysis auxiliary device according to any one of claims 1 to 6, wherein the second item group includes items of health data acquired from a wearable device.   8. The representative determined by the representative determination unit is a representative determined as a target for renting a wearable device for acquiring an evaluation value of each item of the second item group. Health data analysis auxiliary device described in 1.   9. The health data analysis assisting apparatus according to claim 8, wherein the number of representatives determined is larger than the number of wearable devices to be lent, and the wearable devices are lent for a period. The representative determining unit further includes:
Lending in the total period, using as input the total period for which lending is allowed, the period necessary to obtain the evaluation value, and the relationship in which wearable devices can be delivered without causing a time lag between representatives Output the number of wearable devices necessary to complete the process and the rental schedule, or
A schedule for completing the lending in a short period of time according to the progress of the lending, with the relationship between the representatives being able to deliver wearable devices without causing a time lag and the predetermined number of wearable devices to be lent The health data analysis auxiliary device according to claim 9, wherein: The representative determining unit further includes:
A representative from a representative with a low page rank calculated from the relation that can be delivered by inputting the relation that can deliver the wearable device without causing a time lag between representatives and the predetermined number of wearable devices to be lent. The health data analysis auxiliary device according to claim 9, wherein a schedule for completing the lending in a short period of time according to the progress of the lending is output by sequentially lending to the person.   The health data analysis auxiliary device according to any one of claims 1 to 11, wherein the first item group includes items of medical examination data.   The first item group includes items of standard medical examination data, and the second item group includes items of special medical examination data for items other than the items of the standard medical examination data. The health data analysis auxiliary device according to any one of claims 1 to 6, wherein